Knowledge Efficient Federated Continual Learning for Industrial Edge Systems

Abstract

Recent advances in federated learning (FL) primarily focus on addressing inter-client data heterogeneity, implicitly assuming static data within each client. However, this assumption is inadequate for industrial edge systems (IES), which operate in dynamically changing environments and require real-time processing and analysis of voluminous time-series data generated by the Internet of Things. To bridge this gap, we propose MeCo, a novel federated continual learning (FCL) method for IES, designed to avoid forgetting past knowledge while continuously adapting to new task data. MeCo distinguishes itself from traditional FL by effectively addressing both inter-client and intra-client data heterogeneity through a knowledge-efficient strategy. Specifically, it includes: Meta task-invariant knowledge consolidation, which helps capture shared features across tasks to alleviate forgetting; Consistent task-specific knowledge transfer, which allows edge clients to extract relevant knowledge from a server-side knowledge pool, providing a jump-starting for the current task. Experimental results demonstrate that MeCo significantly outperforms other federated and/or continual learning approaches in real-world industrial fault diagnosis, achieving approximately 2% higher Mean Average Accuracy and being 1.74 times more cost-effective in server-to-client communication. These advantages, along with its robust performance in IES, indicate the potential of MeCo for facilitating edge-cloud collaborative learning in the future.